By "cherry picking" I mean a user with permissions to view and modify the workgroup queue will sometimes (even though they shouldn't) manually pickup interactions waiting in queue instead of allowing the ACD engine to deliver them. This prevents the ACD engine from doing its job and can mess up queue statistics.
The definition of a flow out is a call that is neither answered nor abandoned. To narrow down the exact scenario that results in a flow out, first narrow down the time range when it happens. Then look at the interactions within that time range that fit any additional criteria that can help to identify the interaction that triggers a flow out (workgroup queue, agent assigned, etc.). Once you have a small number of interactions to evaluate, first look at the call log for those interactions to see if you see anything obviously different with the number of interactions that match the number of flow outs. Then, if necessary, dig deeper into those interactions in the trace logs. I'd start with the IP log.
If you are not comfortable sifting through interaction reporting and trace log data, I recommend engaging your ININ partner (if indirect) or ININ support (if direct) to assist in tracking down the interactions that are being tagged as flow outs to find out why it is happening.